# Third-party components

Murmello uses NVIDIA and Suno's Parakeet TDT-CTC 110M English model, converted to Core ML by Fluid Inference, and the FluidAudio Swift library. The model library additionally offers NVIDIA's Parakeet TDT 0.6B V2 and V3 models, converted and quantized by Fluid Inference.

- FluidAudio v0.15.6, Apache 2.0: https://github.com/FluidInference/FluidAudio
- Original model, CC BY 4.0: https://huggingface.co/nvidia/parakeet-tdt_ctc-110m
- Core ML conversion: https://huggingface.co/FluidInference/parakeet-tdt-ctc-110m-coreml
- Pinned model revision: `9bc92ead6e8f17eca92a869fd578ae76842b82ba`
- The complete library and model license texts are included in the app's Resources directory.

Optional model downloads (CC BY 4.0 model license):

- NVIDIA Parakeet TDT 0.6B V2: https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2
- V2 Core ML conversion: https://huggingface.co/FluidInference/parakeet-tdt-0.6b-v2-coreml, revision `ee09c569f73759e6d44c9bd16766f477b2b36d39`
- NVIDIA Parakeet TDT 0.6B V3: https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
- V3 Core ML conversion, including the optional INT4 encoder: https://huggingface.co/FluidInference/parakeet-tdt-0.6b-v3-coreml, revision `7dd20fe6b1797d35f5e3307e8b1732d9a178edfe`
- License: https://creativecommons.org/licenses/by/4.0/ (full text included as `Model-LICENSE`). FluidAudio's Apache 2.0 library license is separate from the model license.

The model files are used without modification. Murmello adds audio capture, local inference orchestration, keyboard shortcuts, and text delivery. Credit for the model and its conversion remains with their respective authors.

FluidAudio contains additional components and their notices, including FastCluster and NeMo text processing. The NeMo text-processing license is also included in Resources. See the source distribution and LICENSE for upstream notices. Murmello uses the Parakeet ASR path without speaker identification or cloud transcription. Its optional local language models for writing cleanup are listed below.

The development comparison used Moonshine AI's `two_cities.wav` fixture from version v0.1.5 of https://github.com/moonshine-ai/moonshine. That recording is kept in `.work/downloads` and is not bundled with the application. The sample passage is from Charles Dickens' public-domain *A Tale of Two Cities*.

## Optional local text editor

Murmello uses Qwen3 1.7B Q4_K_M for opt-in local cleanup. The base model is [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B), under Apache 2.0; the license ships as `Qwen-LICENSE`. The [Unsloth GGUF conversion](https://huggingface.co/unsloth/Qwen3-1.7B-GGUF) is pinned to `d7f544eead698dbd1f15126ef60b45a1e1933222`. Model weights are downloaded separately.

The compact editor is [HuggingFaceTB/SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct), under Apache 2.0. Its [official GGUF](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct-GGUF) Q8_0 conversion is pinned to `593b5a2e04c8f3e4ee880263f93e0bd2901ad47f`, SHA-256 `48ab3034d0dd401fbc721eb1df3217902fee7dab9078992d66431f09b7750201`. The upstream [SmolLM license](https://github.com/huggingface/smollm/blob/main/LICENSE) ships as `SmolLM2-LICENSE`. The 386,404,992-byte model is a separate optional download. Only the selected editor runs, and it exits after editing.

The bundled local helper and required dynamic libraries come from [llama.cpp b10809](https://github.com/ggml-org/llama.cpp/releases/tag/b10809), copyright its contributors, under the MIT license. The upstream license is included at `Contents/Resources/llama-LICENSE`. The verified archive is pinned by SHA-256; see `CLEANUP.md` for exact model/runtime metadata. No runtime server is started.